Introduction to Parameterized Complexity M. Pouly Department of Informatics University of Fribourg, Switzerland Internal Seminar June 2006
Outline Introduction & Motivation The Misery of Dr. O The Perspective of Complexity Theories Classical Complexity Theory meets with Criticism 1 Introduction & Motivation 2 3 4
The Misery of Dr. O The Misery of Dr. O The Perspective of Complexity Theories Classical Complexity Theory meets with Criticism The scientist Dr. O has collected a number of data points that support a new theory. But some of these observations are in conflict. Naturally, Dr. O wants to know the minimum set of points that explain the inconsistencies.
The Misery of Dr. O The Misery of Dr. O The Perspective of Complexity Theories Classical Complexity Theory meets with Criticism We model data points by dots and conflics by lines.
The Misery of Dr. O The Misery of Dr. O The Perspective of Complexity Theories Classical Complexity Theory meets with Criticism We model data points by dots and conflics by lines. This is VERTEX COVER!
The Misery of Dr. O The Perspective of Complexity Theories Classical Complexity Theory meets with Criticism The Perspective of Complexity Theories The traditional View: Forget it! VERTEX COVER is NPC. But my data set is rather small 40 or 50 data points... The parameterized View: The best known algorithm has O(kn + (4/3) k k 2 ). This is a good practical algorithm for k 70.
The Misery of Dr. O The Perspective of Complexity Theories Classical Complexity Theory meets with Criticism The Perspective of Complexity Theories The traditional View: Forget it! VERTEX COVER is NPC. But my data set is rather small 40 or 50 data points... The parameterized View: The best known algorithm has O(kn + (4/3) k k 2 ). This is a good practical algorithm for k 70.
The Misery of Dr. O The Perspective of Complexity Theories Classical Complexity Theory meets with Criticism The Perspective of Complexity Theories The traditional View: Forget it! VERTEX COVER is NPC. But my data set is rather small 40 or 50 data points... The parameterized View: The best known algorithm has O(kn + (4/3) k k 2 ). This is a good practical algorithm for k 70.
The Misery of Dr. O The Perspective of Complexity Theories Classical Complexity Theory meets with Criticism Classical Complexity Theory meets with Criticism NP-complete Problems: NPC suggests that exhaustive search is the only approach. NPC discourages the investment of effort. In practice, the size of most instances is naturally bounded. Polynomial Problems: Is O(x 100 ) really practical? Are there algorithms with O(x n ) for n > 4?
The Misery of Dr. O The Perspective of Complexity Theories Classical Complexity Theory meets with Criticism Classical Complexity Theory meets with Criticism NP-complete Problems: NPC suggests that exhaustive search is the only approach. NPC discourages the investment of effort. In practice, the size of most instances is naturally bounded. Polynomial Problems: Is O(x 100 ) really practical? Are there algorithms with O(x n ) for n > 4?
Outline Introduction & Motivation Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice 1 Introduction & Motivation 2 3 4
Setup of Parameterized Complexity Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice Problem input: (x, k) Σ N k is called parameter Parameterized language: Yes instances L Σ N For a fixed k, we denote L k the k-th slice of L: L k = {(x, k) : (x, k) L}
Setup of Parameterized Complexity Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice Problem input: (x, k) Σ N k is called parameter Parameterized language: Yes instances L Σ N For a fixed k, we denote L k the k-th slice of L: L k = {(x, k) : (x, k) L}
Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice Examples of Parameterized Problems Some problems belong to both theories. Example: VERTEX COVER Some problems have a parameterized counterpart: Example: WEIGHTED CNF SAT Instance: Propositional formula X in CNF Parameter: k Question: Does X have a satifying weight k assignment? Other problems only exist in one of the two worlds.
Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice Examples of Parameterized Problems Some problems belong to both theories. Example: VERTEX COVER Some problems have a parameterized counterpart: Example: WEIGHTED CNF SAT Instance: Propositional formula X in CNF Parameter: k Question: Does X have a satifying weight k assignment? Other problems only exist in one of the two worlds.
Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice Examples of Parameterized Problems Some problems belong to both theories. Example: VERTEX COVER Some problems have a parameterized counterpart: Example: WEIGHTED CNF SAT Instance: Propositional formula X in CNF Parameter: k Question: Does X have a satifying weight k assignment? Other problems only exist in one of the two worlds.
Fixed Parameter Tractability I Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice We are interested in languages that are tractable by slice: Definition (FPT = Fixed Parameter Tractability) A parameterized problem L Σ N is fixed parameter tractable if there is an algorithm that correctly decides, for input (x, k) Σ N whether (x, k) L in time f (k) x α, where α is constant and f is an arbitrary function. The relation between good and bad part is defined to be multiplicative. A fundamental property of FPT is that the definition is unchanged if we replace it by f (k) + x α. VERTEX COVER FPT
Fixed Parameter Tractability I Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice We are interested in languages that are tractable by slice: Definition (FPT = Fixed Parameter Tractability) A parameterized problem L Σ N is fixed parameter tractable if there is an algorithm that correctly decides, for input (x, k) Σ N whether (x, k) L in time f (k) x α, where α is constant and f is an arbitrary function. The relation between good and bad part is defined to be multiplicative. A fundamental property of FPT is that the definition is unchanged if we replace it by f (k) + x α. VERTEX COVER FPT
Fixed Parameter Tractability II Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice Definition Let A be a parameterized problem. a) We say that A is uniformly fixed-parameter tractable if there is an algorithm Φ, a constant c, and an arbitrary function f : N N such that the running time of Φ(x, k) is at most f (k) x c, (x, k) A iff Φ(x, k) = 1. b) We say that A is strongly uniformly fixed-parameter tractable if A is uniformly fixed-parameter tractable via some Φ and f such that f is recursive. c) We say that A is nonuniformly fixed-parameter tractable if there is a constant c, a function f : N N and a collection of procedures {Φ k : k N} such that for each k the running time of Φ k is at most f (k) x c and (x, k) A iff Φ k (x, k) = 1.
Fixed Parameter Tractability II Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice Definition Let A be a parameterized problem. a) We say that A is uniformly fixed-parameter tractable if there is an algorithm Φ, a constant c, and an arbitrary function f : N N such that the running time of Φ(x, k) is at most f (k) x c, (x, k) A iff Φ(x, k) = 1. b) We say that A is strongly uniformly fixed-parameter tractable if A is uniformly fixed-parameter tractable via some Φ and f such that f is recursive. c) We say that A is nonuniformly fixed-parameter tractable if there is a constant c, a function f : N N and a collection of procedures {Φ k : k N} such that for each k the running time of Φ k is at most f (k) x c and (x, k) A iff Φ k (x, k) = 1.
Fixed Parameter Tractability II Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice Definition Let A be a parameterized problem. a) We say that A is uniformly fixed-parameter tractable if there is an algorithm Φ, a constant c, and an arbitrary function f : N N such that the running time of Φ(x, k) is at most f (k) x c, (x, k) A iff Φ(x, k) = 1. b) We say that A is strongly uniformly fixed-parameter tractable if A is uniformly fixed-parameter tractable via some Φ and f such that f is recursive. c) We say that A is nonuniformly fixed-parameter tractable if there is a constant c, a function f : N N and a collection of procedures {Φ k : k N} such that for each k the running time of Φ k is at most f (k) x c and (x, k) A iff Φ k (x, k) = 1.
How Parameters arise in Practice Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice Natural Parameters: Frequent in graph theory: Instance: A graph G and a positive integer k Parameter: k Question: Is ω(g) k? ω(g) is a width metric, i.e. Treewidth, Cutwidth,... Implicit Parameters: Frequent in formula problems: Instance: A database d and a query φ. Parameter: φ Question: Is the query relation defined by φ nonempty? Engineering Parameters: Key length in cryptology. Approximation Parameters: Related to the ɛ.
How Parameters arise in Practice Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice Natural Parameters: Frequent in graph theory: Instance: A graph G and a positive integer k Parameter: k Question: Is ω(g) k? ω(g) is a width metric, i.e. Treewidth, Cutwidth,... Implicit Parameters: Frequent in formula problems: Instance: A database d and a query φ. Parameter: φ Question: Is the query relation defined by φ nonempty? Engineering Parameters: Key length in cryptology. Approximation Parameters: Related to the ɛ.
How Parameters arise in Practice Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice Natural Parameters: Frequent in graph theory: Instance: A graph G and a positive integer k Parameter: k Question: Is ω(g) k? ω(g) is a width metric, i.e. Treewidth, Cutwidth,... Implicit Parameters: Frequent in formula problems: Instance: A database d and a query φ. Parameter: φ Question: Is the query relation defined by φ nonempty? Engineering Parameters: Key length in cryptology. Approximation Parameters: Related to the ɛ.
How Parameters arise in Practice Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice Natural Parameters: Frequent in graph theory: Instance: A graph G and a positive integer k Parameter: k Question: Is ω(g) k? ω(g) is a width metric, i.e. Treewidth, Cutwidth,... Implicit Parameters: Frequent in formula problems: Instance: A database d and a query φ. Parameter: φ Question: Is the query relation defined by φ nonempty? Engineering Parameters: Key length in cryptology. Approximation Parameters: Related to the ɛ.
Which k are reasonable? Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice The are currently no theorems that indicate limits on what we might reasonably expect to achive in terms of practical parameter ranges for FPT problems. However, the goodess of a FPT algorithm can be measured by the largest value of k for which f (k) is less than some universal speed constant U. For U = 10 20, we call this value number of Klams that the algorithm is worth.
Which k are reasonable? Setup of Parameterized Complexity Examples of Parameterized Problems Fixed Parameter Tractability How Parameters arise in Practice The are currently no theorems that indicate limits on what we might reasonably expect to achive in terms of practical parameter ranges for FPT problems. However, the goodess of a FPT algorithm can be measured by the largest value of k for which f (k) is less than some universal speed constant U. For U = 10 20, we call this value number of Klams that the algorithm is worth.
Outline Introduction & Motivation Improving Classical Intractability Provable Intractability 1 Introduction & Motivation 2 3 4
Improving Classical Intractability Improving Classical Intractability Provable Intractability Barometer of Intractability: FPT W [1] W [2] W [3]... The W-Hierarchy replaces NPC of classical theory. The higher a problem is located in the W-hierarchy, the more unlike it is to be in FPT. How can we find such classes? We need a complete problem for all W[i]!
Improving Classical Intractability Improving Classical Intractability Provable Intractability Barometer of Intractability: FPT W [1] W [2] W [3]... The W-Hierarchy replaces NPC of classical theory. The higher a problem is located in the W-hierarchy, the more unlike it is to be in FPT. How can we find such classes? We need a complete problem for all W[i]!
Parameterized Reductions Improving Classical Intractability Provable Intractability Definition A parameterized problem L reduces to a parameterized problem L if we can transform an instance (x, k) of L into an instance (x, g(k)) of L in time f (k) x c (f and g are arbitrary computable functions), such that (x, k) is a yes-instance of L if and only if (x, g(k)) is a yes-instance of L.
Improving Classical Intractability Provable Intractability Classical Reductions vs. Parameterized Reductions Are classical poly-reductions also parameterized reductions? CNF IP 2 WEIGHTED CNF WEIGHTED IP 2. CNF 3SAT WEIGHTED CNF WEIGHTED 3SAT. ( Moreover, it is unknown if such a reduction exists.) In general, classical reductions do not have enough structure to induce parameterized reductions.
In Search of Complete Problems Improving Classical Intractability Provable Intractability Classical Completeness: SAT NPC We are looking for a similar theorem for every W[i]. Question: What makes satisfiability difficult? Definition A propositional formula X is called t-normalized if X is of the form conjunction-of-disjunction-... of literals with t alternations. 2-normalized formula corresponds to CNF We belief that SAT for t-normalized formulae is strictly easier than for (t+1)-normalized formulae.
In Search of Complete Problems Improving Classical Intractability Provable Intractability Classical Completeness: SAT NPC We are looking for a similar theorem for every W[i]. Question: What makes satisfiability difficult? Definition A propositional formula X is called t-normalized if X is of the form conjunction-of-disjunction-... of literals with t alternations. 2-normalized formula corresponds to CNF We belief that SAT for t-normalized formulae is strictly easier than for (t+1)-normalized formulae.
Some Completeness Results Improving Classical Intractability Provable Intractability Proposition* t 1 WEIGHTED t-normalized SAT is complete for W[t]. * In reality, the W-classes are defined by use of boolean decision circuits and this proposition rises to a theorem. Theorem (Complexity Barometer) FPT W [1] W [2] W [3]... VERTEX COVER FPT INDEPENDENT SET & CLIQUE are complete for W [1]. Amazing! They are equivalent in classical complexity theory.
Some Completeness Results Improving Classical Intractability Provable Intractability Proposition* t 1 WEIGHTED t-normalized SAT is complete for W[t]. * In reality, the W-classes are defined by use of boolean decision circuits and this proposition rises to a theorem. Theorem (Complexity Barometer) FPT W [1] W [2] W [3]... VERTEX COVER FPT INDEPENDENT SET & CLIQUE are complete for W [1]. Amazing! They are equivalent in classical complexity theory.
Beyond W [t]-completeness Improving Classical Intractability Provable Intractability If we do not impose any bound on the depth of the propositional formula, we arrive at classes hard for t W [t]. WEIGHTED SAT Instance: A boolean formula X. Parameter: k Question: Does X have a weight k satisfying assignment? WEIGHTED CIRCUIT SAT Instance: A decision circuit C. Parameter: k Question: Does C have a weight k satisfying assignment?
Beyond W [t]-completeness Improving Classical Intractability Provable Intractability If we do not impose any bound on the depth of the propositional formula, we arrive at classes hard for t W [t]. WEIGHTED SAT Instance: A boolean formula X. Parameter: k Question: Does X have a weight k satisfying assignment? WEIGHTED CIRCUIT SAT Instance: A decision circuit C. Parameter: k Question: Does C have a weight k satisfying assignment?
Beyond W [t]-completeness Improving Classical Intractability Provable Intractability W [SAT ]: problems reducible to WEIGHTED SAT. W [P]: problems reducible to WEIGHTED CIRCUIT SAT. Theorem FPT W [1] W [2] W [3]... W [SAT ] W [P]. The containment W [SAT ] W [P] seems to be proper but there is no proof for this assumption.
Beyond W [t]-completeness Improving Classical Intractability Provable Intractability W [SAT ]: problems reducible to WEIGHTED SAT. W [P]: problems reducible to WEIGHTED CIRCUIT SAT. Theorem FPT W [1] W [2] W [3]... W [SAT ] W [P]. The containment W [SAT ] W [P] seems to be proper but there is no proof for this assumption.
Beyond W [t]-completeness Improving Classical Intractability Provable Intractability W [SAT ]: problems reducible to WEIGHTED SAT. W [P]: problems reducible to WEIGHTED CIRCUIT SAT. Theorem FPT W [1] W [2] W [3]... W [SAT ] W [P]. The containment W [SAT ] W [P] seems to be proper but there is no proof for this assumption.
Provable Intractability Improving Classical Intractability Provable Intractability Definition (X Classes) Let C be a classical complexity class. We say that a parameterized language L is in XC iff L k C for all k. For instance: the class XP consists of those languages that are slicewise polynomial. Theorem (Complexity Barometer) FPT XP FPT W [1] W [2]... W [SAT ] W [P] XP The connection of the two last theorems shows that at least one inclusion in the W-hierarchy must be strict.
Provable Intractability Improving Classical Intractability Provable Intractability Definition (X Classes) Let C be a classical complexity class. We say that a parameterized language L is in XC iff L k C for all k. For instance: the class XP consists of those languages that are slicewise polynomial. Theorem (Complexity Barometer) FPT XP FPT W [1] W [2]... W [SAT ] W [P] XP The connection of the two last theorems shows that at least one inclusion in the W-hierarchy must be strict.
Outline Introduction & Motivation 1 Introduction & Motivation 2 3 4
Introduction & Motivation Parameterized complexity is surprisingly orthogonal to classical complexity theory. There are no systematic or trivial correlations between the two complexities of a given problem. The W-hierarchy can be regarded as a complexity barometer that measures how close a problem comes to a successful deal with the devil. (Downey & Fellows)